TensorFlow：在Android中的推断之间初始化RNN的状态

We have a working TensorFlow network (graphdef) running on Android and I'm noting that the inference results over time tend to be correlated. That is if label A is returned there will tend to be a stream of A's that follow in time, even when the input data switches to data that should generate B labels. Eventually, the results will switch to B but there seems to be a lag and suggests the RNN is stateful between inference calls. Our network is using RNN/LSTMs.We have a working TensorFlow network (graphdef)